Hello Vercel Team,

I’m looking for technical guidance on designing a production-grade AI agent architecture using Vercel technologies.

I specifically need help understanding the recommended approach for:

• persistent conversation and topic state;
• accurate follow-up understanding;
• reference and entity resolution across multiple turns;
• multi-step planning and tool execution;
• evidence reuse, refresh, and grounding;
• structured reasoning records without exposing private chain-of-thought;
• streaming real execution progress;
• retries, fallbacks, and idempotent response finalization;
• long-running analytical tasks;
• response validation and revision loops;
• observability and evaluation of agent behavior.

I would appreciate guidance on whether these workflows should use the Vercel AI SDK, AI SDK tool loops, Workflow, background functions, queues, or a combination of these systems.

I am mainly looking for practical online guidance, recommended architecture patterns, documentation, code examples, or advice from a Vercel team member experienced with production AI agents.

Thank you.